38 AI-for-Science Papers Yield Six Lessons: From Stable mRNA Formulations to Model Pitfalls
bravo_abad · x · 2026-10-03
Jorge Bravo Abad's weekly briefing links 38 recent AI-for-Science papers (Sept 23–Oct 2) across biology, physics, chemistry, engineering and Earth science, distilling six transferable lessons.
Highlights:
- mRNA stability: AGENT combines high-throughput experiments with Bayesian optimization; six rounds over one month found solid-state lipid nanoparticle formulations retaining 100% bioactivity after 2+ months at 37°C — stability was built into the objective from day one.
- Titanium discovery: jointly optimizing composition and manufacturing conditions achieved high strength with substantial ductility.
- Extrapolation trap: a surgical-expertise model hit AUROC 0.888 on random splits but only 0.571 when tested on surgeons never seen before — change the question, confidence collapses.
Core thesis: the biggest wins come from decisions around the model — what to optimize, what information to reuse, and what evidence to require. Includes a practical lab checklist and 23 short paper summaries.
Related event: Weekly AI for Science Roundup Distills 38 Papers into Six Research Lessons(2 posts)→
More from Research
- ENCODE preprint: neural nets predict super-additive GATA motif epistasis (r=0.55) — anshulkundaje · 2026-11-19
- "Feels Old Man": Resurfaced OpenAI Blackboard Photo on Unified Multimodal Modeling — burny_tech · 2026-11-19
- New TLAPS-Bench blog: formally proving correctness of real-world system specs with TLA+ — tianyin_xu · 2026-11-19
- Astra's proof sidesteps three deep Mathlib results by inventing a new 'capacity' quantity — IgorCarron · 2026-11-19
- Stanford's Kundaje: ChIP-seq resolution limits motif-level deep learning analysis — anshulkundaje · 2026-11-19
- Mahina 4.0 LoRA Adapters Claim Fewer Hallucinations, Devs Seek Stress-Tests — External-Fee-8920 · 2026-11-19